On Feb 9, 2015 10:22 PM, "Chen Fan" <chen.fan.fnst@cn.fujitsu.com> wrote:

On 02/02/2015 10:44 PM, Bjorn Helgaas wrote:

On Sun, Feb 1, 2015 at 10:08 PM, Nishank Trivedi
<nishank.trivedi@netapp.com> wrote:

On 2/1/15 6:13 PM, Chen Fan wrote:

On 02/01/2015 07:01 AM, Kevin O'Connor wrote:

On Wed, Jan 28, 2015 at 04:05:13PM +0800, Chen Fan wrote:

For PCIe device support AER(Advanced Error Reporting), from the
pcie spec 3.0 chapter 6.2.5, ERR_COR, ERR_NONFATAL, and ERR_FATAL
can be forwarded from the secondary interface to the primary interface,
only require the SERR# Enable bit in the Bridge Control register is set.

and at the kernel side, we found only _HPP() method can enable
SERR#, So here we want to turn on this bit.

Signed-off-by: Chen Fan <chen.fan.fnst@cn.fujitsu.com>
---
   src/fw/pciinit.c | 4 ++++
   1 file changed, 4 insertions(+)

diff --git a/src/fw/pciinit.c b/src/fw/pciinit.c
index 34279a4..28ed1af 100644
--- a/src/fw/pciinit.c
+++ b/src/fw/pciinit.c
@@ -310,6 +310,10 @@ static void pci_bios_init_device(struct
pci_device *pci)
       /* enable memory mappings */
       pci_config_maskw(bdf, PCI_COMMAND, 0,
                        PCI_COMMAND_IO | PCI_COMMAND_MEMORY |
PCI_COMMAND_SERR);
+    /* enable SERR# for forwording */
+    if (pci->header_type & PCI_HEADER_TYPE_BRIDGE)
+        pci_config_maskw(bdf, PCI_BRIDGE_CONTROL, 0,
+                         PCI_BRIDGE_CTL_SERR);
   }

Thanks for submitting your patch.

I'm not that familiar with all the details of the PCI specification.
Can you elaborate on what doesn't work if this patch is not present,
or what does work or work better when present?  Is this associated
with a QEMU feature or change?

With this patch,  the aer error inject feature for pcie device in QEMU
will work . the hmp command line is called pcie_aer_inject_error.
a endpoint device propagates error messages to guest that need the
parent bridge type devices enable the SERR# in bridge control reigster
from pcie spec 3.0 chapter 6.2.5, and currently, the register bit is
not turned on, so the command pcie_aer_inject_error is not work.
and if the patch is ok, could you check in this patch?

Patch looks fine to me.
Certain PCI switches, like PEX87xx, provide mechanism to explicitly inject
errors (unlike aer_inject tool where just interrupt is raised). However, if
SERR bit is not enabled in the bridges, error is not percolated to root
port, probably because kernel expects BIOS to enable this bit. CC'ing Bjorn
who can correct if this incorrect understanding.

I think your understanding of PCI_BRIDGE_CTL_SERR is correct.  Linux
doesn't really have formal expectations about how platforms should set
PCI_BRIDGE_CTL_SERR.  It's possible that Linux should be more
proactive about enabling it.  I think it's just an area that nobody
has paid much attention to.

Hi Kevin,

How do you think of this patch?

Thanks,
Chen

Bjorn
.